Forming different-complexity covariance-model subspaces through piecewise-constant spectra for hyperspectral image classification

Abstract

A key factor in classifiers based on the normal (or Gaussian) distribution is the modeling of covariance matrices. When the number of available training pixels is limited, as often is the case in hyperspectral image classification, it is necessary to limit the complexity of these covariance models. An alternative to reducing the complexity uniformly over the whole feature space, is to form orthogonal subspaces and reduce the model complexity within them separately, e.g., forming full-complexity within-class, or interior-class, subspace models, and reduced-complexity exterior-class subspace models. We propose to use subspaces created by forming fewer and wider spectral bands, instead of the more general principal component analysis transform (PCA), in an attempt to exploit a-priori knowledge of the data to create more generalizable subspaces. We investigate the resulting classifiers by studying their performances on four hyperspectral data sets. On each data set, experiments where run using different training set sizes. The results indicate that the classifiers seem to benefit from using this more data-specific approach to forming subspaces. © 2011 Springer-Verlag.